Vehicle velocity forecasting plays a critical role in scheduling the operations of varying systems and devices in a passenger vehicle. This paper first generates a repeated urban driving cycle dataset at a fixed route in the Dallas area, aiming to contribute to the improvement of vehicle energy efficiency for commuting routes. The generated driving cycles are divided into cycle segments based on intersection/stop identification, deceleration and reacceleration identification, and waiting time estimation, which could be used for better evaluating the effectiveness of model localization. Then, a segment-based vehicle velocity forecasting model is developed, where a machine learning model is trained/developed at each segment, using the hidden Markov chain (HMM) model and long short-term memory network (LSTM). To further improve the forecasting accuracy, a localized model selection framework is developed, which can dynamically choose a forecasting model (i.e., HMM or LSTM) for each segment. Results show that (i) the segment-based forecast could improve the forecasting accuracy by up to 24%, compared the whole cycle-based forecast; and (ii) the localized model selection framework could further improve the forecasting accuracy by 6.8%, compared to the segment-based LSTM model. Moreover, the potential of leveraging the stopping location at an intersection to estimate the waiting time is also evaluated in this study.